OctoInformBeta

OpenAI preprint, September 2026 · §4. Conditioning on block types

§4 (setting): block types and conditional translation laws: Finite conditional laws, affine tails with an unconditioned future, and entropy loss at prescribed contraction depths

Conditioning translations on block types.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

A finite address carries both contraction data and translation randomness. Conditioning on block types separates these two features in the finite translation law and in its affine mixture with a future measure.

A weighted affine iterated function system on a finite alphabet 𝒜\mathcal A is a family of similarities φi\varphi_i with strictly positive normalized weights pip_i, as in (1). φi(x)=rix+ti,0<|ri|<1,ti∈ℝ,pi>0,∑i∈𝒜pi=1.\varphi_i(x)=r_i x+t_i,\qquad 0<|r_i|<1,\qquad t_i\in\mathbb R,\qquad p_i>0,\qquad \sum_{i\in\mathcal A}p_i=1.(1)✦ For a word w=(w0,…,wn−1)w=(w_0,\ldots,w_{n-1}), its complete affine map is the pair (rw,tw)(r_w,t_w) of signed slope and translation, its word similarity is φw\varphi_w, and its word probability is pwp_w, defined by (2). φw=φw0∘⋯∘φwn−1,φw(x)=rwx+tw,pw=∏j=0n−1pwj,(r⌀,t⌀)=(1,0).\varphi_w=\varphi_{w_0}\circ\cdots\circ\varphi_{w_{n-1}},\qquad \varphi_w(x)=r_wx+t_w,\qquad p_w=\prod_{j=0}^{n-1}p_{w_j},\qquad (r_{\varnothing},t_{\varnothing})=(1,0).(2)

For positive weights qq on a finite set EE, a tag τ:E→B\tau:E\to B and points a,v∈Ea,v\in E, the tag fiber Fτ(a)F_\tau(a) is the set of points with the same tag as aa, its mass is Mτ(a)M_\tau(a), and the posterior Pτ(a,v)P_\tau(a,v) is the conditional probability of vv given that tag, as in (3). Fτ(a)={v∈E:τ(v)=τ(a)},Mτ(a)=∑v∈Fτ(a)qv,Pτ(a,v)={qv/Mτ(a),v∈Fτ(a),0,v∉Fτ(a).\begin{aligned} F_\tau(a)&=\{v\in E:\tau(v)=\tau(a)\},& M_\tau(a)&=\sum_{v\in F_\tau(a)}q_v,\\ P_\tau(a,v)&= \begin{cases} q_v/M_\tau(a),&v\in F_\tau(a),\\ 0,&v\notin F_\tau(a). \end{cases} \end{aligned}(3) For positive weights qq on a finite set EE, a tag τ:E→B\tau:E\to B and a point a∈Ea\in E, the conditional observable distribution of an observable f:E→ℝf:E\to\mathbb R given the tag of aa is the distribution of ff under the weights Pτ(a,v)P_\tau(a,v) in (3).✦

A finite probability law ν\nu on ℝ\mathbb R is a finite support SνS_\nu with strictly positive point weights summing to one.✦ For strictly positive weights qq summing to one on a finite set EE and an observable f:E→ℝf:E\to\mathbb R, the observable law of ff is the finitely supported probability law on its image whose weight at xx is the fiber mass mf(x)m_f(x) in (4).✦ The Shannon entropy of the weights qq in nats and their observable entropy are also given there, with 0ln⁡0=00\ln0=0. mf(x)=∑a∈Ef(a)=xqa,H(q)=−∑a∈Eqaln⁡qa,H(f;q)=−∑x∈f(E)mf(x)ln⁡mf(x).m_f(x)=\sum_{\substack{a\in E\\f(a)=x}}q_a,\qquad H(q)=-\sum_{a\in E}q_a\ln q_a,\qquad H(f;q)=-\sum_{x\in f(E)}m_f(x)\ln m_f(x).(4) The Shannon entropy H2(ν)H_2(\nu) in bits of a finite probability law ν\nu on ℝ\mathbb R is the real number obtained by dividing its point entropy by ln⁡2\ln2.✦ For a finite probability law ν\nu on ℝ\mathbb R and a scale ρ>0\rho>0, the grid entropy H2,ρ(ν)H_{2,\rho}(\nu) is the entropy in bits of the grid-cell observable of ν\nu, as in (5). H2(ν)=−1ln⁡2∑x∈Sνν{x}ln⁡ν{x},H2,ρ(ν)=H(x↦⌊x/ρ⌋;ν)ln⁡2.H_2(\nu)=-\frac{1}{\ln2} \sum_{x\in S_\nu}\nu\{x\}\ln\nu\{x\},\qquad H_{2,\rho}(\nu)=\frac{H(x\mapsto\lfloor x/\rho\rfloor;\nu)}{\ln2}.(5)

The Claim

Definition 1 (Block types and conditional translation laws). Let the system be as in (1), and let b≥1b\geq1 be an integer. The bb-block system is the weighted affine system on 𝒜b\mathcal A^b whose map at a block uu is φu\varphi_u and whose weight is pup_u.✦ The word type κ(u)\kappa(u) of a block uu in the bb-block system is its vector of symbol counts, defined by (6). κ(u)i=#{k∈{0,…,b−1}:uk=i},i∈𝒜.\kappa(u)_i=\#\{k\in\{0,\ldots,b-1\}:u_k=i\}, \qquad i\in\mathcal A.(6)✦ For a word ww of length nbnb, its block type sequence 𝜿(w)\boldsymbol\kappa(w) is the sequence of count vectors in its consecutive blocks, given by (7). 𝜿(w)j(i)=#{k∈{0,…,b−1}:wjb+k=i},0≤j<n.\boldsymbol\kappa(w)_j(i) =\#\{k\in\{0,\ldots,b-1\}:w_{jb+k}=i\}, \qquad 0\leq j<n.(7)✦ For two blocks u,vu,v with the same word type, their signed ratios satisfy (8). ru=rv.r_u=r_v.(8)✦

Let τ:𝒜→B\tau:\mathcal A\to B be any tag, and let 𝝉(w)=(τ(wj))0≤j<n\boldsymbol\tau(w)=(\tau(w_j))_{0\leq j<n} be the tag sequence of a word.✦ For words w,v∈𝒜nw,v\in\mathcal A^n and a tag τ:𝒜→B\tau:\mathcal A\to B, the conditional word probability πτ(w,v)\pi_\tau(w,v) is the product of the symbol posteriors from (3), with qi=piq_i=p_i, as in (9). πτ(w,v)=∏j=0n−1Pτ(wj,vj).\pi_\tau(w,v)=\prod_{j=0}^{n-1}P_\tau(w_j,v_j).(9)✦ For a tag τ:𝒜→B\tau:\mathcal A\to B and a word w∈𝒜nw\in\mathcal A^n, the conditional translation law ντ,w\nu_{\tau,w} is the distribution of the translation tvt_v conditioned on the tag sequence 𝝉(w)\boldsymbol\tau(w), as defined by (10). ντ,w=∑v∈𝒜n𝝉(v)=𝝉(w)pv∑z∈𝒜n𝝉(z)=𝝉(w)pzδtv.\nu_{\tau,w} =\sum_{\substack{v\in\mathcal A^n\\ \boldsymbol\tau(v)=\boldsymbol\tau(w)}} \frac{p_v}{\displaystyle \sum_{\substack{z\in\mathcal A^n\\ \boldsymbol\tau(z)=\boldsymbol\tau(w)}}p_z} \,\delta_{t_v}.(10)✦ For a tag τ:𝒜→B\tau:\mathcal A\to B, a word w∈𝒜nw\in\mathcal A^n and a Borel measure η\eta on ℝ\mathbb R, the conditional affine tail Tτ,wηT_{\tau,w}\eta is the Borel measure given by the mixture in (11), where (φv)*η(\varphi_v)_*\eta denotes pushforward; the future measure η\eta remains unconditioned. Tτ,wη=∑v∈𝒜nπτ(w,v)(φv)*η.T_{\tau,w}\eta=\sum_{v\in\mathcal A^n} \pi_\tau(w,v)\,(\varphi_v)_*\eta.(11)✦ These definitions also apply to the block system, with its word type as the tag.

The contraction depth λi\lambda_i of a symbol i∈𝒜i\in\mathcal A is the real number given by (12). λi=−log⁡2|ri|.\lambda_i=-\log_2|r_i|.(12)✦ For a word ww over 𝒜\mathcal A, its contraction depth σ(w)\sigma(w) is given by (13). σ(w)=−log⁡2|rw|.\sigma(w)=-\log_2|r_w|.(13)✦ For real parameters A,A′A,A' and an integer n≥0n\geq0, the fine scale ρn\rho_n is the positive real number in (14), and the band start ana_n associated with this scale is the integer in (15). ρn=2−⌈An⌉.\rho_n=2^{-\lceil An\rceil}.(14)✦ an=⌈An⌉−1.a_n=\lceil An\rceil-1.(15)✦ For a tag τ\tau on 𝒜\mathcal A, a word ww over 𝒜\mathcal A and a scale ρ>0\rho>0, the conditional entropy loss Lτ(w;ρ)L_\tau(w;\rho) is the real number measured in nats given by (16). Lτ(w;ρ)=(H2(ντ,w)−H2,ρ(ντ,w))ln⁡2.L_\tau(w;\rho) =\bigl(H_2(\nu_{\tau,w})-H_{2,\rho}(\nu_{\tau,w})\bigr)\ln2.(16)✦ For a tag τ\tau on 𝒜\mathcal A, real parameters A,A′A,A' and an integer n≥0n\geq0, the retained mean entropy loss ℛτ(A,A′;n)\mathcal R_\tau(A,A';n) is the real number obtained by summing the weighted losses over addresses of depth at most A′nA'n, as in (17). ℛτ(A,A′;n)=∑w∈𝒜nσ(w)≤A′npwLτ(w;ρn).\mathcal R_\tau(A,A';n) =\sum_{\substack{w\in\mathcal A^n\\\sigma(w)\leq A'n}} p_w L_\tau(w;\rho_n).(17)✦

In the paper, the result reads: for the self-similar probability measure μ\mu, with lower Hausdorff dimension d=dim⁡Hμ<1d=\dim_H\mu<1, base-two map-walk entropy rate hh and positive base-two Lyapunov exponent χ\chi satisfying dχ<hd\chi<h, choose α>χ\alpha>\chi with dα<hd\alpha<h and an integer b≥1b\geq1 with mlog⁡2(b+1)<b(h−dα)m\log_2(b+1)<b(h-d\alpha), where m=|𝒜|m=|\mathcal A|, and set A=bαA=b\alpha, A′=b(χ+α)/2A'=b(\chi+\alpha)/2 and hb=bh−mlog⁡2(b+1)h_b=bh-m\log_2(b+1), so that bχ<A′<Ab\chi<A'<A and hb−dA>0h_b-dA>0; a bb-letter block WW of independent and identically distributed symbols has count type κ\kappa, signed ratio r(κ)=∏iriκir(\kappa)=\prod_i r_i^{\kappa_i} and uniform conditional law QκQ_\kappa on words of that type; independent types ωh\omega_h have their induced law qq, and, writing Rj=∏h<jr(ωh)R_j=\prod_{h<j}r(\omega_h) and σj=−log⁡2|Rj|\sigma_j=-\log_2|R_j|, a bound DD on one block’s depth gives 0≤σj≤Dj0\leq\sigma_j\leq Dj; the first jj types generate ℱj\mathcal F_j, and conditional on the entire environment ω\omega, independent choices WhW_h with laws QωhQ_{\omega_h} define νω,n\nu_{\omega,n} as the law of ∑h<nRhφWh(0)\sum_{h<n}R_h\varphi_{W_h}(0) and μω\mu_\omega as the corresponding infinite sum law, whose type average is μ\mu; the scale and depth event are ρn=2−⌈An⌉\rho_n=2^{-\lceil An\rceil}, an=⌈An⌉−1a_n=\lceil An\rceil-1 and En={σn≤A′n}E_n=\{\sigma_n\leq A'n\}.

What Is Checked

The constructions are stated over finite weighted affine systems, arbitrary tags on their alphabets and Borel future measures, with real scale parameters. Every nonempty word has signed slope of modulus less than one✦; at each fixed length the word probabilities are strictly positive✦ and sum to one.✦ The normalized weights on every tag fiber sum to one.✦